Comparison of Costello Geomagnetic Activity Index Model and JHU/APL Models for Kp Prediction By: David Marchese Mentors: Douglas Biesecker Christopher Balch
Outline � Background � Kp Prediction � Costello Geomagnetic Activity Index Model � Validation Studies � Research � Results � JHU/APL Models � Conclusions
Kp Index Developed by Julius Bartels � Measure of the maximum disturbances in the � horizontal components of Earth’s magnetic fi eld caused by solar particle radiation Of fi cial index calculated every three hours � using observations from 13 subauroral magnetometer stations
Kp Values � Range from 0 to 9 in a scale of thirds � Kp value of 0 corresponds to the quietest conditions � Kp value of 9 corresponds to the most disturbed conditions � Quasi-logarithmic scale � ap index ranges from 0 to 400 and represents the Kp value converted to a linear scale in nT
Effects of Geomagnetic Storms Disrupt radio communications � � Disrupt GPS navigation Damage transformers and electric power grids � � Degrade satellite instrumentation � Increase satellite drag � Cause aurora Confuse racing pigeons �
NOAA Space Weather Scales NOAA G-Scale based on Kp estimates � from the Boulder-NOAA Magnetometer Warnings issued when Kp values of 4, 5, � 6, and 7 or greater are expected Alerts issued for Kp values of 4, 5, 6, 7, 8, � and 9
NOAA G-Scale
USAF Estimated Kp � Of fi cial Kp index published with signi fi cant time delay � “Nowcast” Kp algorithm provides real-time estimates of Kp � Derived using data from 9 ground-based magnetometers in North America � Calculated by the United States Air Force 55th Space Weather Squadron
Costello Geomagnetic Activity Index Neural network algorithm trained on the � response of Kp to solar wind data Input two hours of data for solar wind � speed, IMF magnitude, and Bz Output running 3-hour Kp every 15 � minutes
Motivation for Research � Space weather forecasters need to know how reliable prediction models are � Several validation studies have been done on the Costello model � Results are not complimentary � Important to determine the reasons for discrepancies
Costello Validation Study 1 Covers the time period from August 17, � 1978 to February 16, 1980 (ISEE-3) Predictions binned to integer values � between 0 and 7 Tends to underpredict high and low Kp � underprediction values overprediction Study performed by members of the Space Environment Center.
Costello Validation Study 2 Covers the time period from 1975-2001 � (IMP-8, Wind, ACE) Of fi cial Kp values obtained by � interpolating between points to match 15 minute time granularity overprediction Tends to overpredict low Kp values and � underpredict high Kp values Correlation coef fi cient = 0.75 � underprediction Study performed by Wing et al.
Research Find the distribution of of fi cial Kp values for a given prediction � Determine if the models perform differently during solar maximum years than � during solar minimum years Compare the performance of the Costello model to the JHU/APL models � Data Set � Supplied Costello prediction data spans � from July 1, 1998 until June 18, 2007 Data gap from May 7, 2005 until April 1, � 2006 Time granularity of 15 minutes � Of fi cial Kp database is essentially � uninterrupted since 1932 Time granularity of 3 hours �
Problem � Time granularity � Model predictions are made approximately every 15 minutes � Of fi cial Kp values are calculated once every 3 hours � Solution � Time-tag each of the of fi cial Kp values at the beginning of the 3 hour interval and fi nd model predictions that are made between 0 and 10 minutes after this time
Costello Validation underprediction underprediction overprediction overprediction Kp bins range from 0+ to 7+ � Figure 1: of fi cial Kp averages for each bin are plotted with error bars one standard � deviation in length Figure 2: the median of fi cial Kp values for each bin are plotted with error bars � showing the upper and lower quartiles
Solar Cycle Dependence � During solar maximum external in fl uences Solar dominate activity Minimum in the magnetosphere � During solar minimum internal dynamics are responsible for Solar Maximum fl uctuations in magnetic fi eld strength
Solar Cycle Dependence (Cont.) Solar Maximum Solar Minimum � Costello model appears to predict low Kp values slightly better during solar maximum years
Forecast Specific Validation Expected Kp of 6 Expected Kp of 7 or greater (G2 storm) (G3 or higher storm) � Figures show the distribution of of fi cial Kp values for Costello predictions corresponding to NOAA warnings
Forecast Specific Validation (Cont.) Expected Kp of 4 Expected Kp of 5 (G1 storm) � Figures show the distribution of of fi cial Kp values for Costello predictions corresponding to NOAA warnings
JHU/APL Models � APL Model 1 � Inputs nowcast Kp and solar wind parameters � Predicts Kp 1 hour ahead � APL Model 2 � Same inputs as APL Model 1 � Predicts Kp 4 hours ahead � APL Model 3 � Inputs solar wind parameters � Predicts Kp 1 hour ahead
APL Model 1 � Inputs nowcast Kp and solar wind parameters � Predicts Kp 1 hour ahead � Correlation coef fi cient = 0.92 overprediction underprediction
APL Model 2 � Inputs nowcast Kp and solar wind parameters � Predicts Kp 4 hours ahead � Correlation coef fi cient = 0.79 overprediction underprediction
APL Model 3 � Inputs solar wind parameters overprediction � Predicts Kp 1 hour ahead � Correlation coef fi cient = 0.84 underprediction
Resolution to Discrepancy? � Interpolation of of fi cial Kp values may lead to skew in Wing’s validations � When no interpolation is Interpolated Interpolated Of fi cial Kp Of fi cial Kp used, APL model tends to overpredict Kp instead of underpredicting � Similar skew may be responsible for discrepancy No in Costello validations No Interpolation Interpolation
APL Model Validations � APL models installed � Code edited to run on a NOAA/SEC computer � Models successfully produce real-time Kp estimates � Real-time data plots were not produced � Modi fi cations to run models off of historical data were not completed
Summary We found that the Costello model tends to overpredict Kp consistently � Model performance may exhibit some solar cycle dependency � Statistical evaluations will have to be performed in order to determine the extent of this � dependency Differences in performance are likely irrelevent for forecasting purposes � Directly comparable validation studies should be carried out to determine if the � JHU/APL models perform signi fi cantly better than the Costello model Time interval, time granularity, and data set used should be identical �
References � Detman, T., and J. A. Joselyn (1999), Real-time Kp predictions from ACE real time solar wind, in Solar Wind Nine , edited by S. R. Habbal et al., AIP Conf. Proc., 471, 729-732. � Wing, S., J. R. Johnson, J. Jen, C.-I. Meng, D. G. Sibeck, K. Bechtold, J. Freeman, K. Costello, M. Balikhin, and K. Takahashi (2005), Kp forecast models, J. Geophys. Res., 110 , A04203, doi:10.1029/2004JA010500. � sd-www.jhuapl.edu/UPOS/ � www.gfz-potsdam.de � www.n3kl.org � www.ngdc.noaa.gov � www.sec.noaa.gov
Acknowledgements and Thanks � NOAA/SEC � Douglas Biesecker � Christopher Balch � JHU/APL � Simon Wing � Janice Scho fi eld
Recommend
More recommend
